[Bugfix] Align vLLM YaRN with Transformers and stop re-scaling max_model_len (#56446)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-09-11 13:01:38 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent a5b714eee4
commit c191787a68
11 changed files with 133 additions and 134 deletions
@@ -252,21 +252,28 @@ class MRotaryEmbedding(RotaryEmbeddingBase):
# YaRN parameters.
*,
scaling_factor: float | None = None,
extrapolation_factor: float = 1,
attn_factor: float = 1,
beta_fast: int = 32,
beta_slow: int = 1,
mscale: float | None = None,
mscale_all_dim: float | None = None,
attention_factor: float | None = None,
truncate: bool = True,
) -> None:
self.scaling_factor = scaling_factor
self.extrapolation_factor = extrapolation_factor
self.attn_factor = attn_factor
self.beta_fast = beta_fast
self.beta_slow = beta_slow
self.truncate = truncate
if self.scaling_factor is not None:
# Get n-d magnitude scaling corrected for interpolation
self.mscale = float(yarn_get_mscale(self.scaling_factor) * attn_factor)
if attention_factor is not None:
self.mscale = float(attention_factor)
elif mscale and mscale_all_dim:
self.mscale = float(
yarn_get_mscale(self.scaling_factor, mscale)
/ yarn_get_mscale(self.scaling_factor, mscale_all_dim)
)
else:
self.mscale = float(yarn_get_mscale(self.scaling_factor))
else:
self.mscale = 1.0